SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

SDXL is a latent diffusion model for text-to-image synthesis that markedly enlarges the Stable Diffusion architecture. Its UNet backbone is three times larger, mostly due to more attention blocks and a larger cross-attention context, and it employs a second text encoder. The authors design multiple novel conditioning schemes, train the model across multiple aspect ratios, and introduce a separate refinement model that improves the visual fidelity of generated samples through a post-hoc image-to-image technique.

The authors demonstrate that SDXL drastically improves performance over previous versions of Stable Diffusion and achieves results competitive with black-box state-of-the-art image generators. In the spirit of open research and transparency around large-model training and evaluation, they release code and model weights publicly, making a strong open text-to-image system broadly available.